…eel inline

The dedicated build stage only ever downloaded a prebuilt release wheel
(no compilation), so the extra stage plus bind-mount indirection was
unnecessary. Instead, strip the transitive triton==3.6.0 pin from
requirements/test/cpu.txt on x86 builds and install the triton-cpu
wheel directly via uv pip install where it's needed (vllm-test-deps
consumers: vllm-dev, vllm-test). vllm-openai needs no changes since it
never installs requirements/test/cpu.txt and already gets triton-cpu
via requirements/cpu.txt.

Signed-off-by: jiang1.li <jiang1.li@intel.com>

…stall

requirements/test/cpu.txt transitively pins a generic PyPI triton (via
xgrammar) that clobbers the triton-cpu wheel from requirements/cpu.txt.
Rather than sed-stripping that pin and manually re-downloading the
triton-cpu wheel by hand-selecting its Python-version tag, just let the
existing install steps fix it up:

- vllm-test already runs `uv pip install dist/*.whl` after test/cpu.txt;
  the vllm wheel's install_requires (generated from requirements/cpu.txt
  by setup.py) already pins the same triton-cpu URL, so uv corrects it
  automatically.
- vllm-dev has no such wheel-install step (it uses `setup.py develop`),
  so reinstall requirements/cpu.txt afterward to restore triton-cpu.
- vllm-openai and vllm-test-deps need no changes.

Signed-off-by: jiang1.li <jiang1.li@intel.com>

….txt

Stripping requirements/test/cpu.txt's generic triton==3.6.0 pin (like
the existing decord/terratorch handling for arm builds) is more direct
than letting it clobber the triton-cpu wheel and then reinstalling
requirements/cpu.txt to fix it back up. Since the pin is never applied,
vllm-dev needs no corrective step at all.

Signed-off-by: jiang1.li <jiang1.li@intel.com>

triton-cpu is a required CPU runtime dependency on x86_64 (see
requirements/cpu.txt) that several CPU kernels quietly fall back
from (spec decode, GDN mamba) via triton_utils.HAS_TRITON. Surface
that expectation explicitly instead of only failing later inside
individual kernels.

Signed-off-by: jiang1.li <jiang1.li@intel.com>

Signed-off-by: jiang1.li <jiang1.li@intel.com>

Signed-off-by: jiang1.li <jiang1.li@intel.com>

…1151/vllm into worktree-triton-cpu-wheel